Transparency makes an AI system’s role, information sources, and limitations visible to the people who use or oversee it. In an enterprise GenAI application, users should know when they are interacting with AI or receiving AI-generated content. They also need enough information to judge whether an output is suitable for their task. Transparency works alongside explainability: disclosing how a system is used is not the same as explaining a particular decision, and LLM limitations make the latter challenging. section 13.3.3 section 13.1
Make transparency part of the application and its operating practices:
- Disclose the AI’s role and identify AI-generated content. Publish policies and relevant limitations so users understand what they are accepting. section 13.3.3
- Cite sources when generated content is reference-based. Let users review and edit outputs before using them, while making their responsibility for the final content clear. section 13.3.3
- Consult model cards and transparency notes for information about model safety processes. Test model versions and assess potential harms rather than treating documentation as proof of safety. section 13.3.3
- Record evaluation criteria, methods, and results so others can verify performance claims. Include bias, fairness, and user experience in that evaluation. section 12.1
Reference-based enterprise assistant
Suppose an assistant drafts an answer from reference content. Its interface can identify the answer as AI-generated, cite the referenced sources, and give the employee a chance to check and edit it before use. If a content filter blocks a response, leaving the employee unable to tell why can erode trust. These measures make the interaction easier to assess; they do not establish that every answer or filtering decision is correct. section 13.3.3 section 13.5.3
Transparency is an ongoing responsible-AI practice, not a one-time notice. Identify and measure harms, mitigate them, and keep operating practices under review. Be candid about potential model biases and limitations, and update disclosures and evaluations as the application changes. section 13.3 section 13.3.4
